Ali Farahani - Department of Computer Engineering, Shahid Bahonar University of Kerman, Kerman, Iran
Hadis Mohseni - Young Researchers Society, Shahid Bahonar University of Kerman, Kerman, Iran

Head pose estimation is a challenging task for many computer vision systems, such as face detection and face recognition. In recent years, several methods have been proposed for head pose estimation. However, most of the proposed methods in this area are based on facial feature detection and location of facial features such as eyes and nose tip is necessary for accurate head pose estimation in them. Although some of these methods have perfect performance for small and medium (about ±45°) head pose angles, uncontrolled conditions such as illumination and occlusion can affect the facial feature detection and pose estimation performance. In this paper, we propose a holistic pose estimation method with low computational cost which is based on linear reconstruction of a face image over its nearest neighbors in training images, regardless of facial feature locations. The proposed method benefits from the assumption that the reconstruction of a face image in an arbitrary pose p from training images in pose p has less error than the reconstruction of that image from training images in other poses. Extensive experiments are conducted on FacePix and CMU-PIE face databases to verify the efficacy and accuracy of the proposed method

head pose estimation, linear reconstruction, nearest neighbors